Did California's Shelter-In-Place Order Work? Early Coronavirus-Related Public Health Effects - IZA DP No. 13160 APRIL 2020 - IZA Institute ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
DISCUSSION PAPER SERIES IZA DP No. 13160 Did California’s Shelter-In-Place Order Work? Early Coronavirus-Related Public Health Effects Andrew I. Friedson Drew McNichols Joseph J. Sabia Dhaval Dave APRIL 2020
DISCUSSION PAPER SERIES IZA DP No. 13160 Did California’s Shelter-In-Place Order Work? Early Coronavirus-Related Public Health Effects Andrew I. Friedson Joseph J. Sabia University of Colorado Denver San Diego State University and IZA Drew McNichols Dhaval Dave University of San Diego-California and Bentley University, IZA and NBER San Diego State University APRIL 2020 Any opinions expressed in this paper are those of the author(s) and not those of IZA. Research published in this series may include views on policy, but IZA takes no institutional policy positions. The IZA research network is committed to the IZA Guiding Principles of Research Integrity. The IZA Institute of Labor Economics is an independent economic research institute that conducts research in labor economics and offers evidence-based policy advice on labor market issues. Supported by the Deutsche Post Foundation, IZA runs the world’s largest network of economists, whose research aims to provide answers to the global labor market challenges of our time. Our key objective is to build bridges between academic research, policymakers and society. IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper should account for its provisional character. A revised version may be available directly from the author. ISSN: 2365-9793 IZA – Institute of Labor Economics Schaumburg-Lippe-Straße 5–9 Phone: +49-228-3894-0 53113 Bonn, Germany Email: publications@iza.org www.iza.org
IZA DP No. 13160 APRIL 2020 ABSTRACT Did California’s Shelter-In-Place Order Work? Early Coronavirus-Related Public Health Effects* On March 19, 2020, California Governor Gavin Newsom issued Executive Order N-33-20 2020, which required all residents of the state of California to shelter in place for all but essential activities such as grocery shopping, retrieving prescriptions from a pharmacy, or caring for relatives. This shelter-in-place order (SIPO), the first such statewide order issued in the United States, was designed to reduce COVID-19 cases and mortality. While the White House Task Force on the Coronavirus has credited the State of California for taking early action to prevent a statewide COVID-19 outbreak, no study has examined the impact of California’s SIPO. Using daily state-level coronavirus data and a synthetic control research design, we find that California’s statewide SIPO reduced COVID-19 cases by 125.5 to 219.7 per 100,000 population by April 20, one month following the order. We further find that California’s SIPO led to as many as 1,661 fewer COVID-19 deaths during the first four weeks following its enactment. Back-of-the-envelope calculations suggest that there were about 400 job losses per life saved during this short-run post-treatment period. JEL Classification: H75, I18 Keywords: coronavirus, COVID-19, shelter in place order, synthetic control Corresponding author: Joseph J. Sabia Department of Economics San Diego State University 5500 Campanile Drive San Diego, CA 92182-4485 USA E-mail: jsabia@sdsu.edu * Sabia acknowledges research support from the Center for Health Economics & Policy Studies (CHEPS) at San Diego State University, including grant support received from the Charles Koch Foundation and the Troesh Family Foundation. We thank Kyutaro Matsuzawa, Isaac Baumann, Cameron Milani, Alicia Marquez, James Curry, and Samuel Safford for excellent research assistance on this project. We also thank Donald Kenkel, Daniel Hamermesh, David Powell, Sebastian Tello-Trillo, Jason Fletcher, Ben Hansen, Sarah Stith, Laura Argys, Susan Averett, David Neumark, Daniel Rees, and Nicolas Ziebarth for helpful comments and suggestions that helped to improve this paper.
1. Motivation "I simply do not know if our aggressive actions early on ... have had the intended effect … I certainly am hoping and praying that that is the case. We still need the data to confirm that.” - Grant Colfax, San Francisco Director of Public Health, March 31, 2020 The 2020 U.S. coronavirus outbreak is one of the most serious public health challenges in American history. The U.S. has seen more reported COVID-19 cases between January 1, 2020 and April 16, 2020 than reported polio cases between 1910 and 2010 (Ochmann and Roser 2020). The transmission of COVID-19 is presently believed to occur largely through exposure to respiratory droplets, usually emitted during coughing, sneezing, or nose-blowing (Centers for Disease Control and Prevention 2020a), but may also occur through natural respiratory expulsion (Fineberg 2020). The primary strategies suggested by governments worldwide to reduce the spread of COVID-19 have been frequent handwashing, mask-wearing in public, and social distancing (Australian Government Department of Health 2020; Public Health England 2020; Public Health Agency of Canada 2020; White House 2020). In contrast to many countries throughout the world, many of the authorities used to combat public health threats in the United States rest not with the Federal government, but largely with state and local officials. The primary state and local policy strategy to prevent the spread of coronavirus is the enactment of shelter-in-place orders (SIPOs), sometimes called “stay at home” orders. The State of California was at the forefront of issuing comprehensive SIPOs as the COVID-19 crisis unfolded nationwide. On March 17 and 18, 2020, 12 California counties 1
and the City of Berkeley adopted SIPOs. 1 And on the evening of March 19, Governor Gavin Newsom announced the nation’s first statewide SIPO, earning immediate attention in the national news media. California Executive Order N-33-20 2020 stated: “To protect public health, I as State Public Health Officer and Director of the California Department of Public Health order all individuals living in the State of California to stay home or at their place of residence except as needed to maintain continuity of operations of the federal critical infrastructure sectors…” (California Executive Order N-33-20 2020). This statewide executive order required all California residents to remain in their homes for all but essential activities such as purchasing food or medicine, caring for others, exercise, or traveling for employment deemed essential. While grocery stores, pharmacies, restaurants providing takeout or delivery service, and other essential businesses were permitted to remain open, most other non-essential businesses were ordered closed. 2 In addition, residents were advised to continue to maintain a six-foot distance with non-household members with whom they come in contact and public gatherings of non-household members were strongly discouraged. Violations of the SIPO were subject to a $1,000 fine and up to 6 months of imprisonment (Allday 2020), though enforcement most often occurred through social pressure and warnings for 1 On March 16, Los Angeles Mayor Eric Garcetti implemented a non-essential business closure order. The seven counties that adopted a SIPO on March 17 were: Alameda, Contra Costa, Mendocino, San Francisco, San Mateo, Santa Clara, and Santa Cruz. The five counties that adopted a SIPO on March 18 were: Monterey, San Benito, Solano, Sonoma, and Ventura. 2 California’s statewide SIPO was soon followed by the closing of many public parks or beaches (California Department of Parks and Recreation 2020), driven in part by public outrage over the surge in beach parties and picnics immediately following the executive order (Kopetman 2020). 2
first offenses. State SIPOs appear to have had increased social distancing (Abouk and Heydari 2020; Fry 2020), with non-essential travel falling 40 to 60 percent following California’s SIPO (COVID-19 Community Mobility Report). In issuing the SIPO, Governor Newsom implored residents of California “to meet this moment and flatten the curve together” (Romero 2020). Thus, an important policy rationale was not simply to curb the pandemic’s growth in California, but also to delay its peak, allowing the state additional time to obtain the necessary ventilators, hospital beds, and medical staff in place to meet the surge in demand for services among those who test positive (Baker and Fink 2020; Greenstone and Nigam 2020; Ranney et al. 2020; Tsai et al. 2020; Hicks and Marsh 2020). The White House Coronavirus Task Force singled out two states, including California, for their early adoption of mandatory social distancing policies: “[Dr. Deborah] Birx said the experiences of California and Washington give her hope that other states can keep the coronavirus under control through social distancing. That’s because they moved quickly to contain the early clusters of coronavirus by closing schools, urging people to work from home, banning large gatherings and taking other measures now familiar to most Americans.” (The Associated Press, March 31, 2020) Following the adoption of California’s SIPO, 39 additional states and the District of Columbia enacted similar statewide SIPOs between March 20, 2020 and April 9, 2020. And by early April, over 90 countries worldwide, representing half of the world’s population, requested or ordered their citizens to stay at home (Sandford 2020). 3
There are a number of important reasons to single out California’s SIPO in studying its effects on COVID-19 infection and mortality. First, California was the first state to implement a statewide order, an important policy feature designed to prevent within-state spillovers of infectious disease across neighboring counties, cities, and townships with heterogeneous local orders. Second, California ranks 12 out of 51 states (including the District of Columbia) in population per square mile (World Population Review) and 2 out of 51 in percent living in urbanized areas and urban clusters (Iowa Community Indicators Program). 3 The spread of coronavirus is exacerbated by increased population density, which generates greater opportunities for transmission among frequently interacting individuals (Centers for Disease Control and Prevention 2020b). This makes California a location with a high potential for rapid spread of the disease. Third, California implemented its SIPO at a unique time: early in the cycle of an outbreak, when the rate of growth in observed coronavirus case rates was quite low relative to other states that ultimately implemented SIPOs. The average daily growth in confirmed coronavirus cases in the four days prior to SIPO enactment was 19.1 percent in California. No other SIPO-implementing state in the upper 25th percentile of the state population density or urbanicity distribution had an average daily coronavirus growth rate lower than California during this period. As a comparison, New York, which had the highest per-capita case rate in the nation throughout much of the outbreak, had an average daily coronavirus case growth rate of 39.5 percent in the four days prior to its enactment of a SIPO. In this context, California may provide a cleaner natural experiment in relation to later adopters, while also informing the efficacy of the decision to enact a SIPO early in the outbreak cycle. Finally, at this early stage of assessing the 3 In 2019, the population density in California was 253.7 persons per square mile; 95.0 percent of residents lived in an urbanized area or urban cluster. 4
short-run public health effects of SIPOs, California has more days of post-treatment data than any other state, critical for analyzing the public health benefits of an epidemic with potentially exponential short-run contagion and lagged mortality effects. While there is speculation by the White House Coronavirus Task Force that SIPOs have been effective at “bending the case curve,” there is little rigorous empirical analysis of this question. This study is the first to examine early public health effects of a statewide SIPO. Using 29 days of state-level data on confirmed COVID-19 cases and COVID-19-related deaths, this study provides evidence on the public health effects of California’s SIPO during the crucial early weeks of the policy. Estimates from our preferred synthetic control models show that California’s SIPO led to a 125.5 to 219.7 per 100,000 population reduction in COVID-19 cases and a 1.9 to 4.2 per 100,000 population reduction in COVID-19-related deaths per 100,000 during the first month following its enactment. These findings are robust to the selection of observables (i.e. pre-treatment COVID-19 rates, population density, urbanicity, and other COVID-related policies) to generate the weights to construct our counterfactual. We find that the number of cases averted and lives saved were much larger in the second and third weeks following the SIPO’s adoption, consistent with growing public health benefits over the period during which the outbreak was exponentially growing. Back-of-the-envelope calculations suggest that there were approximately 8 to 14 job losses per coronavirus case averted and 421 to 917 job losses per life saved during this short-run post-SIPO window in California. 2. Background on Coronavirus In December 2019, an outbreak of a new zoonotic coronavirus, SARS-CoV-2, which causes the disease COVID-19, was detected in Wuhan, Hubei Province, China (World Health Organization 2020a). The symptoms of COVID-19 resemble that of influenza, including fever, 5
cough, and shortness of breath (Centers for Disease Control and Prevention 2020c). However, severe respiratory symptoms generate a mortality rate that may be an order of magnitude greater than that of the common flu (World Health Organization 2020b, House Oversight and Reform Committee 2020). Deaths from COVID-19-related illness are heavily concentrated among the elderly and those with underlying health conditions such as respiratory ailments, heart disease, diabetes, or high blood pressure (Wu and McGoogan 2020; Verity et al. 2020; CDC Covid-19 Response Team 2020). The substantial epidemiological risk from this new coronavirus is exacerbated by (i) a high degree of contagion (Thackeray 2020; Cascella et al. 2020), (ii) the possibility that transmission may occur among those who are asymptomatic (Centers for Disease Control and Prevention 2020a; Mandavilli 2020, Whitehead 2020, Li et al. 2020), and (iii) lack of adequate testing (Cohen 2020, Shear et al. 2020, Rabin et al. 2020, Desiderio and Levine 2020). In the main, transmission of COVID-19 occurs via respiratory droplets, usually emitted during coughing, sneezing, or nose-blowing (Centers for Disease Control and Prevention 2020a, World Health Organization 2020c & 2020d). There is also some evidence that the disease may be easily transmitted through normal breathing function in common air space (Fineberg 2020). The incubation period for COVID-19 is 2 to 14 days, during which time transmission may also be possible (Centers for Disease Control and Prevention 2020b, Li et al. 2020). U.S. pharmaceutical companies, in partnership with the public sector, have begun clinical trials for a COVID-19 vaccine (White House 2020, Food and Drug Administration 2020). However, a longer-run medical solution such as a vaccine is unlikely to solve short-run capacity constraints faced by hospitals, which have faced a sharp increase in demand for medical services among coronavirus patients (Secon 2020, Seewer 2020, Kaste 2020, Marquez and Moghe 2020). 6
Fears over shortages of ventilators (Baker and Fink 2020, Ranney et al. 2020), intensive care unit beds (Tsai et al. 2020, Hicks and Marsh 2020), and protective personal equipment (James 2020, Petras and Loehrke 2020, Ranney et al. 2020, World Health Organization 2020e) have instead led policymakers to seek strategies to “flatten the curve” of infections, reducing the peak number of infections and spreading its incidence over time (Meredith 2020). The World Health Organization (WHO) recommended that individuals frequently wash hands, avoid shaking hands with others, self-isolate when presenting flu-like symptoms, and avoid all unnecessary contact with non-household members through the practice of social distancing (World Health Organization 2020b). The adoption of shelter-in-place orders is the dominant policy strategy to induce such distancing. SIPOs can affect COVID-19 cases and deaths through a number of channels. As public health officials and epidemiologists expect, reducing exposure of uninfected individuals to those who are infected via social distancing should reduce the transmission rate of the virus. However, there are other indirect channels through which SIPOs could impact reported cases and deaths. SIPOs may affect testing and coronavirus-related mortality because infected individuals who are unaware of their infection choose to stay at home rather than seek out testing or other medical care. This may be because of fears of contagion at medical facilities or because of a desire to adhere to a “civic duty” to shelter in place. In addition, as SIPOs may necessitate an increase in public resources spent on patrol and enforcement, they may reduce resources available for public health services. On the other hand, as SIPOs are expected to reduce the demand for “non- essential” medical procedures such as well care visits or elective surgery, they may free up resources for those most at risk for coronavirus-related mortality. 7
3. Data and Methodology 3.1 Data In order to examine the “first stage” compliance with the California SIPO, we utilize publicly available data from SafeGraph Inc. 4 For each state (and county) on each day SafeGraph provides a shelter-in-place index, based on the percent of individuals staying at home during the day. The metric is calculated from spatial data generated using anonymous cell phone pings. First, each cell phone is assigned a “home” (or 153m by 153m square) based on a common nighttime location over a baseline period. SafeGraph then calculates the percent staying at home, i.e. the fraction of cell phones in a geographic unit (state, county, etc.) that do not leave the “home” for any given day. 5 The shelter-in-place index is the percentage point change in the number of cell phones staying at home relative to the baseline of February 6, 2020 through February 12, 2020. 6 To examine the short-run public health effects of the statewide order, we utilize a panel of state-specific daily counts of COVID-19 cases and deaths from March 12, 2020 through April 20, 2020. These data are collected by the Centers for Disease Control and Prevention (CDC) and made public by the Kaiser Family Foundation. 7 Over this period, there were a total of 778,328 confirmed COVID-19 cases in the United States, 4.3 percent (33,862) of which were in California, and 37,372 coronavirus-related deaths, 3.2 percent (1,223) of which were in California. 4 Data and detailed descriptions of variable construction are available at: https://www.safegraph.com/dashboard/covid19-shelter-in-place 5 SafeGraph makes adjustments for small geographic units which are not relevant for a state-level analysis. 6 So a value of 25 is for the shelter in place index could represent an increase from of 12% of phones staying at home at baseline to 37% of phones staying at home (37-12=25). 7 See data available here: https://www.nytimes.com/interactive/2020/us/coronavirus-us-cases.html 8
Figure 1 shows state-specific trends in cumulative coronavirus rates per 100,000 population in each state. Daily cases can be calculated as the slope of this cumulative case distribution. Over the period under study, the average coronavirus case rate in California was 30.8 per 100,000 population, and its growth rate from March 12 to April 9 ranked 6th lowest among the 50 states and the District of Columbia. In addition, California had the lowest case growth rate among the top 25th percentile of most highly urbanized states. California also had the lowest case rate among the top 25th percentile of most highly densely populated states. Figure 2 shows the trends in cumulative coronavirus-related death rates per 100,000 population. Between March 12 and April 20, the COVID-19 death rate in the United States grew by 8.56 deaths per 100,000 population, while it grew by less than one half as much (3.08 per 100,000 population) in California. New York, New Jersey, Washington, Louisiana, and Massachusetts had among the highest rates of coronavirus-related deaths at the end of the sample period. In terms of the death rate, the growth in the COVID-19 death rate in California ranked lowest among the top 25th percentile of most highly urbanized states. Figure 3 shows trends in coronavirus case rates for each county in the state of California during the sample period. As noted above, prior to the statewide SIPO on March 19, 12 counties and the City of Los Angeles adopted SIPOs. Early adopting counties are shaded black and all other counties are shaded gray. We find that the growth rate in coronavirus cases from March 12 through April 20 in early adopting California counties was about 9 percent. This compares to an average growth rate of 11 percent in the remaining counties in the state who were bound by the statewide order. 8 8 Appendix Figure 1 shows trends in coronavirus mortality rates for each county in California during the sample period. Due to many low population counties with zero deaths for much of the sample period it is difficult to draw reliable inference with regards to county level SIPOs and mortality. 9
3.2 Synthetic Control Design We use the synthetic control method introduced by Abadie et al. (2010) to infer the causal impact of SIPOs on the number of confirmed coronavirus cases and number of COVID- 19-related deaths per 100,000 population in California. This method uses data on pre-treatment COVID-19 case (or mortality) rates and observable characteristics of states that may influence the spread of the virus (i.e. population density, urbanicity, emergency decrees for major disaster area, travel restrictions, and the number of tests conducted). 9 The synthetic control approach generates a counterfactual designed to capture how coronavirus cases would have evolved in California in the absence of its SIPO. Our chief outcomes of interest, and ℎ , measure the cumulative number of confirmed coronavirus cases and number of coronavirus-related deaths per 100,000 population in state i at day t. We estimate the unobserved counterfactual (“synthetic California”) as a weighted linear combination of states included in a donor pool. The weights are chosen so as to generate a synthetic state that is as similar as possible to California on key observables. Given the importance of the selection (i) of states to be included in the donor pool, and (ii) observable characteristics on which to closely match California to its synthetic counterpart, it is 9 Population density is measured as the ratio of state population to land area in square miles and is available from U.S. Census (see: https://www.census.gov/geographies.html). Urbanicity is measured as percentage of total population living in urban areas and is available from Iowa Community Indicators Program (see: https://www.icip.iastate.edu/tables/population/urban-pct-states). Emergency major disaster declarations are measured as a state with disaster that exceeds the response capabilities of the state and local governments, and long- term recovery assistance is needed and are available from the Federal Emergency Management Agency (see: https://www.fema.gov/disasters). Travel restrictions are measured as states that restrict residents from traveling to other states and/or states that restrict residents of other states from entering the state. COVID-19 tests are measured as the natural log of total coronavirus tests reported by each state. These data are available from COVID Tracking Project (see: https://covidtracking.com). 10
incumbent on researchers to offer a theoretical defense of these choices and to explore the sensitivity of estimated policy impacts to these choices (Ferman 2019). Our primary donor pool is comprised of a total of 42 state and the District of Columbia: 10 states that had never enacted a SIPO during our sample period, and 32 states and the District of Columbia that adopted a SIPO at least 5 days after California did so. We select this five-day period because it is the median incubation period of COVID-19 (Lauer et al. 2020), and thus gives the case data from California sufficient time to reflect underlying changes in the transmission rate before any of the donor states implement their own SIPOs. Coupled with information on the date at which each state implemented a SIPO (Appendix Table 1), our decision rule eliminates the following states from our donor pool: Illinois, New Jersey, New York, Connecticut, Louisiana, Oregon, and Washington. 10 Moreover, as shown in Figures 1 and 2, several of these states (New Jersey, New York, and Washington) appeared to be on very different pre-treatment case and mortality trends than the majority of states, including California. To test the sensitivity of our results to the inclusion of these 43 states (including the District of Columbia) in the donor pool, we re-estimate our synthetic control models without eliminating pre-March 24 SIPO-adopters from the donor pool. Our main findings are largely unchanged (see column 1 of Appendix Table 2). One limitation to our preferred donor pool is that by including later-adopting SIPO states as potential donors, the synthetic control is contaminated on some post-treatment days. Thus, to the extent that later enacted SIPOs have taken effect, our estimated treatment effects will be biased towards zero and lower-bound estimates. As one approach to address this concern, we 10 Ohio implemented its statewide SIPO at 11:59pm on March 23, 2020. We code this order as effective on March 24, 2020. However, if we exclude Ohio from our donor pool, we obtain a pattern of results quantitatively similar to those reported above. Moreover, if we allow all early adopters into our donor pool (see column 1 of Appendix Table 2), the results are also unchanged. 11
select an alternate donor pool that includes states that had never adopted a SIPO or had adopted a SIPO but had 4 or fewer days of post-treatment data, meaning that likely at least half of all individuals infected post-SIPO had not yet reached the median incubation period (Lauer et al. 2020). However, this restriction to the donor pool has its own limitation: only 10 states had not adopted SIPOs by April 9 (Arkansas, Iowa, Kentucky, Massachusetts, Nebraska, North Dakota, Oklahoma, South Dakota, Utah, and Wyoming), and just 2 were late adopters that fit this selection criteria (Missouri and South Carolina). 11 Many of these non-adopting and late- adopting states do not share urbanicity, population density, or other policy-related characteristics of California making them potentially bad donors to the synthetic control group. Thus, selecting the eligible donor states for the analyses to follow has an inherent tradeoff. We can select from a broader set of donor states that allow for a better synthetic control match to California on observables, but partially contaminates our control group with treated days biasing the estimated policy effect towards zero, or, we can select from a more limited set of donor states that is uncontaminated by treatment but that may be poorly matched to California on observables leading to a synthetic control group that is potentially a poor counterfactual for California. We will present results under both choices. The former approach yields policy effects that suggest California’s SIPO was effective at slowing the spread of COVID-19 (and is presented in Tables 1 and 4); the latter approach also yields policy effects that suggest California’s SIPO was effective (and is presented in columns 2 and 3 of Appendix Table 2), with results that are larger in magnitude (consistent with the prior approach generating a lower bound estimate) but also less precise. It is our opinion that a better synthetic control match that yields a 11 Among these states, Massachusetts had issued a “stay at home advisory.” 12
downward biased estimate is more reliable than an estimate from a potentially poorly matched control group, a point to which we will return below. With regard to the choice of observables used to select our synthetic control from among the donor states, we take several approaches. First, an important objective of this research design is to generate common pre-treatment trends in coronavirus cases (or deaths). Our primary strategy to achieve this objective is to generate a synthetic control that closely approximates coronavirus case (and mortality) rates in California in each of seven (7) pre-treatment days (March 12 through March 18). We also experimented with requiring pre-treatment matches up to two weeks (14 days) prior to California’s SIPO and our results are unchanged. We estimate the unobserved counterfactual COVID-19 case rate for California on pre-treatment day by ∑ ∗ , where is the weight assigned to donor state . The analogous counterfactual death rate is ∑ ∗ ℎ . The estimated weights are chosen to minimize the absolute difference between = , and ∑ ∗ and for all pre-treatment days, as well as the absolute difference between ℎ = , and ∑ ∗ ℎ . Then, the per-day treatment effect αt is estimated as = = , − ∑ ∗ for t ∈ r [March 19, April 9], where = [ , ℎ ]. The average treatment effect is then the average over the post-treatment window. Choosing a counterfactual based only on pre-treatment outcomes has a number of advantages. Given a long enough pre-treatment window, matching on covariates and matching on outcomes will produce similar results. Choosing not to match on covariates can eliminate concerns of ‘p-hacking’ and reduce bias of our estimate (Botosaru and Ferman, 2017). However, a limitation of matching all pre-treatment days to construct synthetic California is that there may be other observables that may legitimately and strongly influence the path of an 13
outbreak such as population density, urbanicity, weather, state travel restrictions, or a declaration of major disaster. As shown by Kaul et.al (2018), matching on all periods of pre-treatment outcomes renders all covariates irrelevant in the prediction of the outcome. Thus, we take alternate approaches to generating a synthetic control that allow for these relevant observables to influence estimation of synthetic weights. First, we generate weights by matching on coronavirus case (or death) rates on three pre-treatment days (March 13, March 15, and March 18) as well as state population density and a state urbanicity index. These latter factors play an important role in the spread of infectious disease across communities due to increased crowding (Florida 2020). New York City, the most densely populated city in the U.S., for example, experienced one of the most dramatic increases in coronavirus outbreaks and has more coronavirus cases per capita than the nation of Italy (Rosenthal 2020). Second, we generate synthetic weights using March 13, March 15, and March 18 coronavirus case (or death) rates and several state policies designed to affect the outbreak: whether an emergency major disaster declaration was in effect for the state and whether the state had imposed travel restrictions (i.e. guidelines for residents and/or visitors to quarantine themselves for two weeks following arrival from other states). Third, we consider the role of state COVID-19 testing rates. One reason why we might see changes in cases is because testing resources have changed. As of March 13, only 15,000 tests had been conducted in the U.S. To address the low testing rate in the U.S., the Food and Drug Administration approved a new COVID-19 test from the pharmaceutical company Roche (Arnold 2020). In the following days, states including Delaware, New York, Massachusetts and Texas, began implementing drive-up testing sites, which made access to testing more available (Yancey-Bragg 2020). Despite these improvements in accessibility, many testing delays 14
persisted due to laboratory capacity constraints (Brown and Court 2020). Coronavirus-related deaths are less likely to be affected by this selection into testing. Nevertheless, we take two approaches to explore the issue of testing capacity. For our first approach, we explore if California’s SIPO affected testing rates, using synthetic control methods described and the number of tests as the outcome variable. For our second approach, we return to our main analysis of case (and death) rates and generate a synthetic California using March 13, 15 and 18 pre-treatment case (and death) rates as well as the average rate of testing per 100,000 population. To conduct hypothesis tests, we take two approaches. First, we conduct placebo tests following the method suggested by Abadie et al. (2010). We apply the same synthetic control method described above for each donor state. For donor state j, we reclassify the state as a treatment state, and select a donor pool based on all donor states except j, and add California to the donor pool. We then use the same matching criteria stated above to generate a synthetic control ℎ for each donor state j as if treatment were assigned March 19. We compare the pre-treatment and post-treatment mean squared prediction error (MSPE) for each state, calculating the MSPE ratio as follows: ∑ =3/19( − ℎ ) 2 = 3/18 (1) ∑ =3/12( − ℎ )2 The MSPE measures a relative goodness of fit for each state’s synthetic control based on estimated coronavirus cases (or deaths). A high MSPE ratio can be interpreted as poor post- treatment fit relative to pre-treatment fit. The ranking of the treated states relative to the placebo states then provides a permutation-based p-value. The p-value relies directly on the number of states in the donor pool. With 43 states in the donor pool, California’s MSPE ratio must rank at 15
least 2nd to receive a p-value below .05 and at least 4th to receive a p-value below .1. A p-value of less than .01 is not attainable given the number of donor states. 12 The p-value generated for the “standard” synthetic control model generates information on post-treatment fit based on the entire post-treatment period following the March 19 California order. However, there are some important reasons why we might want to calculate MSPE over particular post-treatment windows. For example, differences in the California “gap” (the difference between California’s coronavirus rate and synthetic California’s coronavirus rate) is expected to be much smaller during the incubation period (median of 5.1 days, 97.5th percentile of 11.5 days) following the SIPO relative to the post-incubation period when symptoms would have been present and infectious disease transmission more efficient. Moreover, given that the effects of social distancing may have important multiplier effects over time (Florida 2020), we could expect the “gap” to increase exponentially following treatment. For these reasons, we provide a range of permutation p-values based on several post-treatment windows: (i) the entire post-treatment period, (ii) the post-SIPO period after the median incubation length (March 24 and later), and (iii) the post-SIPO period after which 97.5 percent of those infected on the SIPO implementation date are expected to experience symptoms (March 30 and later). Second, we conduct statistical inference by estimating wild bootstrap cluster standard errors (Cameron et al. 2008; Cameron and Miller 2015) using difference-in-differences regressions. We draw data from California and donor states that received positive synthetic 12 The null hypothesis utilizing these p-values is that there is no effect in each period, not that the average policy effect is zero. It is possible to have a policy effect with an average of zero in the post-treatment period, but a large MSPE if the effect of the policy changes sign in the post-period. For this reason, it is important to make judgements based on both the MSPE and the general pattern of results. 16
weights in the construction of synthetic California and estimate a difference-in-differences model of the following form: ln (Case Rateit) = β0 + β1SIPOit + β2Xit + γs + τt + μit (2) where SIPOit is an indicator for the day that California adopts a shelter-in-place order, Xit is a set of state-level, day-varying controls that include indicators for whether the state enacted a statewide non-essential business closure order (that falls short of a shelter-in-place order), whether the state enacted a targeted shelter-in-place order that covers only older individuals over age 65 or those with underlying health conditions, whether the state had enacted travel-related restrictions, whether a major disaster declaration had been issued for the state, the average temperature (in degrees Celsius) at all weather stations in the state, and an indicator for whether any weather station in the state reported measurable precipitation. In addition, we control for state fixed effects (γs) and day fixed effects (τt), identifying our treatment effect (β1) via within- state variation in the enactment of a SIPO. We also estimate equation (2) where we replace Case Rateit with Death Rateit, but given low counts of deaths during the earlier days of our analysis period, we experiment with count data models such as Poisson and negative binomial models. We first estimate equation (2) using an unweighted linear regression, giving equal weight to California and each positively synthetically-weighted donor state. Then, we use the synthetic weights (between 0 and 1) generated from the synthetic control model to weight each state and then assign California a weight of 1. In all cases, to conduct statistical inference, we generate standard errors via wild cluster bootstrapping given the small number of clusters (states), which includes a single treatment state (Cameron et al. 2008; Cameron and Miller 2015). 17
4. Results Our main findings on the effects of California’s statewide SIPO can be found in Figures 4 through 18 and Tables 1 through 6. Discussion of county-specific COVID-19 SIPOs are discussed in Section 4.4 and in Tables 7 and 8. 4.1 “Stay at home” Behavior and California Statewide SIPO Figure 4 presents trends in the shelter-in-place index for both California and its synthetic control. The index reflects the percent of the population in a given state that stays at home all day relative to a baseline, derived from anonymized cell phone geotagging (SafeGraph, Inc.). We assign weights to the synthetic control based on close matches in the shelter-in-place index for each of the 7 days prior to CA’s statewide SIPO. 13 Trends in social distancing are expectedly positive over the entire analysis period as awareness of COVID-19 was proliferating and the benefits of social distancing were emphasized through public health advisories and guidelines. The synthetically-generated counterfactual tracks California nearly identically prior to the SIPO, with trends markedly diverging only after California issued its statewide order on March 19. Estimates of the average daily treatment effect indicate that the percent of individuals remaining at home throughout the day increased by 2.1 percent in California relative to its synthetic control, over the entire post-SIPO period. The effect is larger over the initial post-treatment window up to March 28, indicating an average 3.3 percent increase per day in individuals sheltering at home in California relative to the control, and suggestive of a rapid run-up in social distancing. 13 As the baseline for the index is the same for all states (week ending February 12, 2020), and the pre-SIPO trends are nearly identical, the treatment effect can be interpreted as the increase in the percent of households in CA who are staying at home relative to its counterfactual in the post-SIPO period. 18
Focusing on this window wherein CA experienced the largest gains in social distancing, the effect is statistically significant with a one-sided permutation-based p-value of 0.023. As more donor states issued their own shelter-in-place orders, notably by the end of March, and gains for California decelerated, trends narrowed somewhat between California and its synthetic control. These analyses underscore two points. First, they provide some supporting evidence that individuals in California complied with the shelter-in-place order, and that the SIPO effectively and rapidly reduced social mobility in California, an effect which was sustained over the analysis period. Second, as pressure for social distancing intensified, the health effects that we estimate below capture the direct effect of California’s SIPO on contagion by forcing certain forms of economic activity to stop as well as the compounding effects of the SIPO accelerating social distancing behavior during the early period of the coronavirus outbreak cycle. 4.2 COVID-19 Confirmed Cases and California Statewide SIPO Figure 5a shows trends in California’s confirmed COVID-19 cumulative cases from March 12, 2020 through April 20, 2020. Case rates rose fairly linearly over the period from March 12 through March 25 from .6 to 8 per 100,000 population before beginning more exponential growth, reaching 85.6 per 100,000 on April 20, 2020. Our first synthetic control assigns weights based on close matches in pre-SIPO case rates on each of the 7 pre-treatment days. We find that a total of 10 states contributed approximately two-thirds of the weight for the estimated counterfactual. These include Colorado, the District of Columbia, Hawaii, Maine, Massachusetts, Michigan, Minnesota, New Mexico, Rhode Island and South Dakota (Figure 5b). 14 This estimated synthetic control serves as our counterfactual for 14 The highest synthetic weights were assigned to Colorado (11%), the District of Columbia (3%), Maine (10%), Massachusetts (15%), and South Dakota (18%). 19
coronavirus case trends that would have unfolded in California in the absence of the SIPO enactment. Figure 5a shows that COVID-19 case rate trends in California and synthetic California are nearly identical in the pre-March 19 period. During the first three (3) days following treatment, which capture the first half of the coronavirus incubation period, coronavirus case rates remain quite similar in California and the synthetic control state. However, beginning on March 23 and accelerating after March 25, the rate of growth in California’s coronavirus cases was substantially lower in California relative to the synthetic control. By the final day of case data in our analysis (April 20), we find that there were 125.5 fewer coronavirus cases per 100,000 population, which translates to 49,636 fewer total cases on that day. This pattern of findings suggests that the public health benefits of the SIPO grew over the near-month long period of analysis. These public health benefits also capture any potential compounding effects due to California responding early during the initial phase of its outbreak cycle. 15 In column (1) of Table 1, we show estimates of the average daily treatment effect for the full post-SIPO period from March 19 through April 9 (Panel I), for the post-treatment period beginning March 23, which accounts for the median coronavirus incubation period (Panel II), and for the post-SIPO period beginning March 30, the 97th percentile of the coronavirus incubation period (Panel III). For the full post-treatment period, we estimate that the enactment of the SIPO is associated with a 42.02 decline in average coronavirus cases per 100,000 15 If we employ the synthetic weights from Model 1 (Table 1) to compute a weighted average date of SIPO enactment among the states in the control group and assign never adopters a pseudo-treatment date that is one day after the end of the analysis period, then, using this measure, California issued its SIPO on average 13 days prior to synthetic California. 20
population. Over the near month-long window, this suggests approximately 49,636 COVID-19 cases averted. 16 The daily decline in coronavirus cases is not constant over the post-treatment window. The estimated public health benefits accelerate in the days following enactment, consistent with an exponential growth in contagion that was averted from the SIPO’s enactment. Following March 22, the enactment of the SIPO led to a 47.7 decline in average coronavirus cases per 100,000 population while in the post-March 30 period, the average cumulative case reduction that we attribute to California’s SIPO is 61.0 per 100,000 population. Together, this evidence is consistent with the hypothesis that the public health benefits of California’s SIPO grew exponentially in the short-run. To conduct statistical inference of the above estimates, we turn to permutation-based p- value calculations. In Figure 5c, we show results from our placebo tests, which give rise to the rank-based p-values we generate. In each case, the permutation-based p-value was .045 using the standard ranking generated by the pre-post MSPE ratio, and .045 using a one-tailed pre-post MSPE ratio test. An important concern in synthetic control analysis is how robust findings are to the choice of observable controls and donor states. Figure 6a shows results from a synthetic control approach where we only assure pre-treatment matches on COVID-19 cases on three days (March 13, March 15, and March 18), but also augment our matching strategy by assuring similarity on the share of the state population residing in urban areas. According to the 2010 Census, 95.0 percent of the population lived in urbanized areas or urban clusters. In Appendix Table 3, we 16 To take a more conservative approach, we also consider an alternate strategy where we force coronavirus cases during the period from March 19 through March 23 to be comparable to the donor states. This imposes a null effect of the SIPO on cases due to lagged effect from exposure to symptoms during the incubation period. This exercise, shown in Appendix Figure 4, yields similar results as our main model. 21
compare urbanicity rates in California (column 1) with all other states (column 2), and synthetic California (column 3). We find that urbanicity rates are most similar to California with the synthetic counterfactual. Despite generating a very different set of positively weighted donor states from the approach undertaken in Figure 5a, the estimated treatment effect is quite similar, as is the pre- treatment match. 17 There is a tight pre-treatment fit, with an exponential divergence following the post-SIPO initial virus incubation period. In column (2) of Table 1, we show that over the full post-treatment period, the California SIPO led to a 59.3 per 100,000 population decline in average cumulative coronavirus cases. Again, the effect sizes were larger in later days of the post-treatment period (Panels II and III), consistent with exponential growth in public health benefits to those in California. The permutation-based p-value from placebo tests (Figure 6c) is .023 in all panels. Given that urbanicity rates may not fully capture the degree of crowding, in Figures 7a through 7c, we show results from a synthetic control model where we match not only on urbanicity, but also population density, increasing the likelihood that we more heavily weight those donor states with higher risk of coronavirus contagion. 18 Our findings, shown in Figures 7a-7c and in column (3) of Table 1, are quite similar to those in column (2) of Table 1. To ensure that coronavirus testing is not biasing estimates of the effect of California’s SIPO on confirmed cases, we next select weights for donor states by also matching on the average rate of testing over the sample period. 19 The results in Appendix Figures 2a through 2c 17 In Appendix Table 4, we show synthetic control estimates where we limit the donor states to the continental United States, thus eliminating Alaska and Hawaii as potential donors. The pattern of results is qualitatively similar to that shown in Table 2. 18 Appendix Table 3 shows the quality of the synthetic control on these observables. 19 Appendix Table 3 shows means of testing rates for California and its synthetic control. 22
provide some support for the hypothesis that testing for the coronavirus may have increased following the SIPO, suggesting that estimated case reductions may be biased toward zero. After controlling for average coronavirus testing rates, we find that the California SIPO led to an average reduction of 57.8 COVID-19 cases per 100,000 population. By April 20, we estimate 68,157 COVID-19 cases averted (Figures 8a-8c and column 4 of Table 1). In column (5), we explore whether our estimated treatment effect is sensitive to the addition of COVID-19-related policy controls, including major disaster declarations and state travel restrictions (see Appendix Table 3). The pattern of results shown in Figures 9a through 9c are quite similar to those shown in the previous figures. We find that the California SIPO led to an average 60.0 case per 100,000 population reduction in COVID-19 cases. By April 20, the cumulative number of coronavirus cases averted was approximately 66,020 (column 5 of Table 1). One important question about our prior analyses is whether we are identifying the effect of a SIPO or just voluntary social distancing that is correlated with the adoption of a SIPO and with coronavirus cases. In column (6) of Table 1, we match on pre-treatment growth in rates of staying-at-home using SafeGraph data as well as three days of pre-treatment coronavirus cases. Here, we estimate that over the near-month long pretreatment period, California’s SIPO was associated with a nearly 87,000 decline in the total number of COVID-19 cases. Figures 10a-10c show these results. Could our donor states be too different from synthetic California in experiencing community spread of the coronavirus? Different states experienced coronavirus outbreak at different times. In the pre-treatment period, California had experienced nearly 252 cases of COVID-19 on March 12, which suggests that outbreak and community spread was well 23
underway. To ensure that the donor pool for synthetic California had also experienced community spread, in Table 2, we limit our donor pool to states that had experienced at least 10 (column 1), 50 (column 2), or 100 cases (column 3) of coronavirus. The results in Table 2 and Figure 11 continue to show strong evidence that California’s SIPO was associated with a reduction in COVID-19 cases. It is important to recall that each of the above estimates is based on a synthetic control from a set of donor states that are partially treated in the latter part of the post-period, and as such are a lower bound for the true effect of the policy. 20 In columns (2) and (3) of Appendix Table 2, we examine the sensitivity of estimated policy impacts to the use of never-adopting or later- adopting (after April 5) states as controls. We find that California’s statewide SIPO reduced average cumulative case rates by between 44.4 and 75.2 cases per 100,000 relative to synthetic California. These results are larger in magnitude, but also less precise than the estimates when we allow partially treated states in the control group. Appendix Figure 7 presents the corresponding trends in cumulative cases for California and synthetic California. While these results are validating, we note here the qualification that the set of donor states that receive positive weights among non-adopters include Massachusetts (38 percent), Oklahoma (22.4 percent), Utah (22.3 percent), North Dakota (14 percent), and Wyoming (3.2 percent). Appendix Table 5 shows that the synthetic California estimated from never adopters (columns 2 and 3) and never or late adopters (columns 4 and 5) differ on population density and urbanicity. Given the important role of social interactions in the spread of 20 In Appendix Figures 5 and 6, we trim the post-treatment window in order to limit the effects of later SIPOs enacted by states within the synthetic control. Specifically, Appendix Figure 5 trims the post-treatment period to March 28, which is 5 days (median incubation period) after the first treatment date for any donor state, ensuring that the trend in cases exhibited by synthetic CA is not due to a SIPO enacted within the control group. Appendix Figure 6 slightly extends the post-treatment period to April 3, ensuring that the window ends before the median incubation period plus lag to ARDS (13 days). Both figures show consistent evidence that the statewide SIPO in CA flattened its outbreak trajectory and reduced the average daily COVID-19 cases. 24
the coronavirus, this makes it possible that the synthetic California generated from a much narrow set of never (or later) adopting donor states were on a different longer-run coronavirus contagion trend, and thus may serve as a poor counterfactual for California. 21 Table 3 presents difference-in-differences estimates of the effect of California’s SIPO on the natural log of confirmed COVID-19 cases. In each case, the sample is comprised of California and each donor state that received a positive weight in the synthetic control exercise described by each column, which follows the set of tests from Table 3. Panel I shows estimates from unweighted regressions, while Panel II uses synthetic weights (which sum to 1) for the donor states and a weight of 1.0 for California. All models include the full set of controls described below equation (2) above. Standard errors are generated via wild cluster (state) bootstrap. Our findings provide consistent evidence that the California SIPO had substantial negative effects on coronavirus cases. Unweighted estimates in Panel I show that the enactment of California’s SIPO was associated with a 50.3 to 58.9 percent decline in the coronavirus rate. When we weight using synthetic weights (Panel II), the effect sizes range from approximately 35.5 to 53.7 percent declines. 22 In Panel III of Table 3, we focus on regressions weighted by synthetic weights and decompose the average treatment effect based on the approximate interquartile range of the incubation window following enactment of the SIPO. Consistent with the synthetic control estimates, we find that the effectiveness of the SIPO grows larger following the virus’s incubation period and the period from time until first symptom until respiratory failure. 21 This may also help explain why the estimated effects for deaths in Panel II of Appendix Table 2 are weak and do not attain statistical significance at conventional levels. The longer lag from infection to incubation to ARDS limits the quality of the synthetic control match when drawing from the smaller pool of never or late adopters. 22 Given that the outcome is the log of the caseload, the percent decline for model 1 is [e-0.439- 1]*100 = 35.5%. 25
Together, the findings in Tables 1 through 3 provide compelling evidence that California’s first-in-the-nation SIPO generated important public health benefits in preventing the spread of the coronavirus during the first three weeks of enactment. 23 Next, we explore whether these case declines generated improvements in mortality rates. 4.3 COVID-19 Deaths and California Statewide SIPO In Figures 12a through 12c, we show estimates of the effect of California’s SIPO on coronavirus-related mortality. Figure 12a shows the exponential rise in cumulative COVID-19- related deaths in California, from 0.01 per 100,000 population (4 deaths) on March 12 to 3.09 per 100,000 population (1,223 deaths) on April 20. When we generate a synthetic control using all pre-treatment days to generate matches on COVID-related mortality rates, we find strong evidence of a pre-March 19 common trend for California and synthetic California. On March 25, 6 days following enactment of the SIPO, we find that the rate of increase in mortality begins falling in California as compared to the estimated counterfactual, with a gap that widens exponentially over time, getting particularly large nearly two weeks after the policy’s adoption on March 31, 2020. This longer lagged effect on mortality relative to cases is to be expected given the incubation period from exposure to symptoms and time from first symptoms to acute respiratory distress syndrome (ARDS), the latter of which may take up to 8 days (Wang et al. 2020). Moreover, a more conservative synthetic control approach, shown in Appendix Figures 8a through 8c, which requires pre-treatment mortality rate matches for every period through the sum 23 In Appendix Figures 3a through 3c, we show synthetic control results when we use daily case rates rather than cumulative case rates. This isolates a different local average treatment effect; that is, the effect of a SIPO on the rate of change in the change in cumulative coronavirus cases. The results continue to provide evidence that California’s SIPO reduced coronavirus-related mortality. 26
You can also read